bool getLastGTIDinOplog(GTID* gtid) {
Client::ReadContext ctx(rsoplog);
// TODO: Should this be using rsOplogDetails, verifying non-null?
NamespaceDetails *d = nsdetails(rsoplog);
shared_ptr<Cursor> c( BasicCursor::make(d, -1) );
if (c->ok()) {
*gtid = getGTIDFromOplogEntry(c->current());
return true;
}
return false;
}
bool getLastGTIDinOplog(GTID* gtid) {
LOCK_REASON(lockReason, "repl: looking up last GTID in oplog");
Client::ReadContext ctx(rsoplog, lockReason);
// TODO: Should this be using rsOplogDetails, verifying non-null?
Collection *cl = getCollection(rsoplog);
shared_ptr<Cursor> c( Cursor::make(cl, -1) );
if (c->ok()) {
*gtid = getGTIDFromOplogEntry(c->current());
return true;
}
return false;
}
// returns number of seconds to sleep, if any
uint32_t BackgroundSync::produce() {
// normally msgCheckNewState gets called periodically, but in a single node repl set
// there are no heartbeat threads, so we do it here to be sure. this is relevant if the
// singleton member has done a stepDown() and needs to come back up.
if (theReplSet->config().members.size() == 1 &&
theReplSet->myConfig().potentiallyHot()) {
Manager* mgr = theReplSet->mgr;
// When would mgr be null? During replsettest'ing, in which case we should
// fall through and actually apply ops as if we were a real secondary.
if (mgr) {
mgr->send(boost::bind(&Manager::msgCheckNewState, theReplSet->mgr));
// There should never be ops to sync in a 1-member set, anyway
return 1;
}
}
OplogReader r(true /* doHandshake */);
// find a target to sync from the last op time written
getOplogReader(r);
// no server found
GTID lastGTIDFetched = theReplSet->gtidManager->getLiveState();
{
boost::unique_lock<boost::mutex> lock(_mutex);
if (_currentSyncTarget == NULL) {
// if there is no one to sync from
return 1; //sleep one second
}
}
r.tailingQueryGTE(rsoplog, lastGTIDFetched);
// if target cut connections between connecting and querying (for
// example, because it stepped down) we might not have a cursor
if (!r.haveCursor()) {
return 0;
}
try {
// this method may actually run rollback, yes, the name is bad
if (isRollbackRequired(r)) {
// sleep 2 seconds and try again. (The 2 is arbitrary).
// If we are not fatal, then we will keep trying to sync
// from another machine
return 2;
}
}
catch (RollbackOplogException& re){
// we attempted a rollback and failed, we must go fatal.
log() << "Caught a RollbackOplogException during rollback, going fatal" << rsLog;
theReplSet->fatal();
return 2; // 2 is arbitrary, if we are going fatal, we are done
}
while (!_opSyncShouldExit) {
while (!_opSyncShouldExit) {
{
// check if we should bail out
boost::unique_lock<boost::mutex> lck(_mutex);
if (!_opSyncShouldRun) {
return 0;
}
}
if (!r.moreInCurrentBatch()) {
// check to see if we have a request to sync
// from a specific target. If so, get out so that
// we can restart the act of syncing and
// do so from the correct target
if (theReplSet->gotForceSync()) {
return 0;
}
verify(!theReplSet->isPrimary());
if (shouldChangeSyncTarget()) {
return 0;
}
//record time for each getmore
{
TimerHolder batchTimer(&getmoreReplStats);
r.more();
}
//increment
networkByteStats.increment(r.currentBatchMessageSize());
}
if (!r.more()) {
break;
}
// This is the operation we have received from the target
// that we must put in our oplog with an applied field of false
BSONObj o = r.nextSafe().getOwned();
opsReadStats.increment();
LOG(3) << "replicating " << o.toString(false, true) << " from " << _currentSyncTarget->fullName() << endl;
uint64_t ts = o["ts"]._numberLong();
// now that we have the element in o, let's check
// if there a delay is required (via slaveDelay) before
// writing it to the oplog
if (theReplSet->myConfig().slaveDelay > 0) {
handleSlaveDelay(ts);
{
boost::unique_lock<boost::mutex> lck(_mutex);
if (!_opSyncShouldRun) {
break;
}
}
}
{
Timer timer;
bool bigTxn = false;
{
Client::Transaction transaction(DB_SERIALIZABLE);
replicateFullTransactionToOplog(o, r, &bigTxn);
// we are operating as a secondary. We don't have to fsync
transaction.commit(DB_TXN_NOSYNC);
}
{
GTID currEntry = getGTIDFromOplogEntry(o);
uint64_t lastHash = o["h"].numberLong();
boost::unique_lock<boost::mutex> lock(_mutex);
// update counters
theReplSet->gtidManager->noteGTIDAdded(currEntry, ts, lastHash);
// notify applier thread that data exists
if (_deque.size() == 0) {
_queueCond.notify_all();
}
_deque.push_back(o);
bufferCountGauge.increment();
bufferSizeGauge.increment(o.objsize());
// this is a flow control mechanism, with bad numbers
// hard coded for now just to get something going.
// If the opSync thread notices that we have over 20000
// transactions in the queue, it waits until we get below
// 10000. This is where we wait if we get too high
// Once we have spilling of transactions working, this
// logic will need to be redone
if (_deque.size() > 20000) {
_queueCond.wait(lock);
}
if (bigTxn) {
// if we have a large transaction, we don't want
// to let it pile up. We want to process it immedietely
// before processing anything else.
while (_deque.size() > 0) {
_queueDone.wait(lock);
}
}
}
}
} // end while
if (shouldChangeSyncTarget()) {
return 0;
}
r.tailCheck();
if( !r.haveCursor() ) {
LOG(1) << "replSet end opSync pass" << rsLog;
return 0;
}
// looping back is ok because this is a tailable cursor
}
return 0;
}
void BackgroundSync::applyOpsFromOplog() {
GTID lastLiveGTID;
GTID lastUnappliedGTID;
while (1) {
try {
BSONObj curr;
{
boost::unique_lock<boost::mutex> lck(_mutex);
// wait until we know an item has been produced
while (_deque.size() == 0 && !_applierShouldExit) {
_queueDone.notify_all();
_queueCond.wait(lck);
}
if (_deque.size() == 0 && _applierShouldExit) {
return;
}
curr = _deque.front();
}
GTID currEntry = getGTIDFromOplogEntry(curr);
theReplSet->gtidManager->noteApplyingGTID(currEntry);
// we must do applyTransactionFromOplog in a loop
// because once we have called noteApplyingGTID, we must
// continue until we are successful in applying the transaction.
for (uint32_t numTries = 0; numTries <= 100; numTries++) {
try {
numTries++;
TimerHolder timer(&applyBatchStats);
applyTransactionFromOplog(curr);
opsAppliedStats.increment();
break;
}
catch (std::exception &e) {
log() << "exception during applying transaction from oplog: " << e.what() << endl;
log() << "oplog entry: " << curr.str() << endl;
if (numTries == 100) {
// something is really wrong if we fail 100 times, let's abort
::abort();
}
sleepsecs(1);
}
}
LOG(3) << "applied " << curr.toString(false, true) << endl;
theReplSet->gtidManager->noteGTIDApplied(currEntry);
{
boost::unique_lock<boost::mutex> lck(_mutex);
dassert(_deque.size() > 0);
_deque.pop_front();
bufferCountGauge.increment(-1);
bufferSizeGauge.increment(-curr.objsize());
// this is a flow control mechanism, with bad numbers
// hard coded for now just to get something going.
// If the opSync thread notices that we have over 20000
// transactions in the queue, it waits until we get below
// 10000. This is where we signal that we have gotten there
// Once we have spilling of transactions working, this
// logic will need to be redone
if (_deque.size() == 10000) {
_queueCond.notify_all();
}
}
}
catch (DBException& e) {
sethbmsg(str::stream() << "db exception in producer on applier thread: " << e.toString());
sleepsecs(2);
}
catch (std::exception& e2) {
sethbmsg(str::stream() << "exception in producer on applier thread: " << e2.what());
sleepsecs(2);
}
}
}
